CONSUMPTION, INCOME AND

COINTECRATION

HSIANG-LINC HAN and MASAO OCAKI

ABSTRACT

This paper examines the long-run relationship of consumption and income in the United States. It was
shown by King et al. (1991) in a version of the permanent income hypothesis that log levels of consump-
tion and income are cointegrated with a known cointegrating vector (1, -1)‘. This implies a restriction that
the cointegrating vector, which eliminates the stochastic trends, also eliminates the deterministic trends
arising from the drift terms of difference stationary variables. Two different methodologies are used to test
this deterministic cointegration restriction in this study. The first methodology is the canonical cointegrat-
ing regression proposed by Park (1992). The second methodology is testing stationarity of the difference
of log consumption and log income, by assuming the cointegrating vector to be (1, -1)‘. Two test statistics
for the null hypothesis of stationarity are employed. One is the G@,q) test proposed by Park and Choi
(1988) and the other is proposed by Kwiatkowski et al. (1992). The results of the study indicate that the
deterministic cointegration restriction cannot be rejected. JEL: E21, C32

I. INTRODUCTION

This paper examines the long-run relationship of consumption and income in the United
States by applying the concept of cointegration. Cointegration was first introduced by
Engle and Granger (1987). Since then there are many empirical studies applying the con-
cept to test validities of economic theories. King et al. (1991) presented a version of the
permanent income hypothesis (PIH) with a constant intertemporal elasticity of substitution

Direct all correspondence to: Hsiang-Ling Han, Economics Division, Babson College, Babson Park, MA
02157-0310; Masao Ogaki, Department of Economics, The Ohio State University, Columbus, OH 45321.

International Review of Economics and Finance, 6(2): 107-l 17 Copyright 0 1997 by JAI Press Inc.
ISSN: 1059-0560 All rights of reproduction in any form reserved.

107
108 HAN and OCAKI

which implies that the difference of log consumption and log income is stationary. It means
that log consumption and log income are cointegrated with a known cointegrating vector
(1, -1)’ according to Engle and Granger (1987). It also means that the deterministic coin-
tegration restriction is satisfied if both log consumption and log income are modeled with
deterministic trends that arise from the drift terms of difference stationary variable as well
as stochastic trends. The deterministic cointegration restriction states that the cointegrating
vector, which eliminates the stochastic trends, also eliminates the deterministic trends.
Other earlier empirical studies, for example, Campbell (1987), Engie and Granger
(1987), Phillips and Ouliaries (1988), focused on Zeveis of consumption and income rather
then log levels of consumption and income to examine the percent income hypothesis.
As shown by Cochrane and Sbordone (1988), there is another version of the PIH which
implies levels of consumption and income are cointegrated. Assume that C(t) is the level of
consumption and Y(t) is the level of income. If both C(t) and Y(t) are difference stationary
with a drift term, then Cochrane and Sbordone’s version of the PIH implies the cointegrat-
ing vector (1, -1)’ should eliminate both stochastic and deterministic trends of C(t) and
Y(t). Since C(t) - Y(t) is stationary while Y(t) grows, this version of the PIH also implies
C(r) - Y(t)
that the saving rate, declines as Y(r) grows. However, this result contradicts
Y(r) ’
with the finding by Kuznet (1946) that the saving rate is stable in the United States in the
long-run. Thus, the version of the PIH proposed by King et al. (1991) is more appropriate
to examine the long-run relationship of consumption and income.
Previous empirical studies on cointegration of consumption and income, including King
et al. (1991), only focused on stochastic trends of these variables. King et al. (1991) did not
find any evidence against the cointegration restriction of the PIH.’ The primary objective
of this study is to test the cointegration restriction on deterministic trends to verify the
validity of the version of the PIH proposed by King et al. (1991).
The deterministic cointegration restriction is implied by many economic models. There
are at least two reasons why testing this restriction is important. First, a test for the deter-
ministic cointegration restriction is important as a specification test for models implying
cointegration. Since it is difficult to test existence of stochastic trends (see Cochrane, 1988;
Christian0 & Eichenbaum, 1989; Campbell & Perron, 1991), it would be difficult to test
any restriction focusing only on stochastic trends. Because it is arguably easier to test exist-
ence of dete~inistic trends, tests including det~~inistic trends and stochastic trends will
be usefut in comparing competing models. Second, it is important to impose the determin-
istic restriction on estimators of cointegrating vector because of efficiency gains for these
estimators, which was demonstrated by West (1988), Hansen (1992), and Park (1992). If
the deterministic cointegration restriction cannot be rejected (e.g., Cooley & Ogaki, 1996),
then the cointegrating vector should be estimated with the deterministic restriction
imposed. On the other hand, if empirical data do not support this restriction, the cointegrat-
ing vector must be estimated from cointegrating stochastic trends. For example, the
deterministic cointegration restriction was rejected by Ogaki and Park (1996) and Ogaki
(1992). The results lead these authors to explore economic specifications that imply only
stochastic cointegration but not deterministic ~ointegration in their models.
There are two different approaches to test the dete~inistic cointegration restriction. The
first approach is to add time polynomials into canonical cointegrating regressions and then
to test the null hypothesis that if coefficients of these time polynomials are zero. When the
Consu~pf;on, Income and Cojnfegratjon 109

cointegrating vector has been estimated and has been incorporated in the canonical cointe-
grating regression, if all coefficients of the added time polynomials are zero, the
deterministic cointegration restriction is satisfied; if all coefficients of the added time poly-
nomials are zero, with the exception of the first order time polynomi~, stochastic
cointegration is the case. The second approach is to test the residuals of a linear combina-
tion of variables with an assumed cointegrating vector. If the series of residuals is
stationary around a fixed level, then the deterministic cointegration is satisfied.
The rest of this paper is organized as follows: Section II discusses economic implications
of cointegration and defines the deterministic cointegration restriction; Section III presents
three different economet~c procedures to test the dete~inistic cointegration restriction-
one is canonical cointegration regressions and the other two are tests for stationarity of
residuals; Section IV reports the empirical results; Section V contains our concluding
remarks.

II. ECONOMIC IMPLI~TIONS OF COINTEG~TION

Let X(t) be a two dimensional difference stationary process, i.e., Z(1) process. X(t) - X(t -
1) = u + v(t) for t 2 1, where p, the drift term, is a two dimensional vector of non-zero real
numbers; and v(f), the error term, is stationary with mean zero and each component of v(t)
has a positive long-run variance. This assumption rules out dete~inistic trends of orders
higher than or equal to quadratic. The initial value of Xi(O), i = 1,2, is an arbitrary random
variable. Then,

x,(t) = pit + x;(t), (1)

where X!(I) = Xi(O) -F x: = , vi(z). Since X:(r) - Xi’(t - 1) = Vi(t), X:(r) is an Z(1) process
without a drift. Relation (1) decomposes a difference stationary process with a drift term
into a deterministic trend arising from the drift term and a difference stationary process
without a drift term.
Suppose that X,(t) and X2(r) are cointegrated with a cointegrating vector (1, -p)‘, the lin-
ear combination of X,(t) - pX2(t) should be stationary. Equation (1) implies

X,(t) - PX2(t) =(cl1 - P&N + [X;(r) - BX; tel. (2)

Therefore, the stationarity of X,(t) - /3X2(t) requires not only Xy (t) - PXi {t) is stationary,
but also pi - &t2 = 0. The first requirement is called stochastic cointegration which
requires only the stochastic components of both series are cointegrated. It is possible that
Xl(t) and X2(t) are stochastically cointegrated while l.11- puz is not zero, therefore X,(t) -
PX2(t) is not stationary. The second requirement is called deterministic cointegration
restriction, ul - fil.12= 0, which implies that (1, -p)’ not only eliminates the stochastic trend
but also the deterministic trend. The deterministic cointegration restriction was not tested
in most of applied economics papers using the concept of cointegration. However, the fol-
lowing two methodologies allow us to test both stochastic and deterministic cointegration
restrictions. The first methodology is a multivariate approach directly applying the concept
110 HAN and OCAKI

of cointegration. The second is an univ~ate approach to test the null hypothesis of level
stationarity for residuals of a linear combination of variables, assuming the cointegrating
vector is known.

fff. ECONOMETRfC PROCEDURES

A. Cointegra ting
Canonical Regression 03i')
Park (1992) first introduced the CCR procedure that involves a data transformation
which uses only the stationary components of a cointegrating model. A cointegrating rela-
tionship supported by the ~ointegrating model would remain unchanged after such a data
tr~sfo~atio~. The CCR transformation makes the error term in a cointegrating model
uncorrelated at the zero frequency with regressors. Therefore, the CCR procedure yields
asymptotically efficient estimators and provides asymptotic chi-square tests that are free
from nuisance parameters.
Suppose that Xt(t} and X*(b) are cointegrated with a cointegrating vector (1, -p)’ and the
error term in the cointegrated system is I. Assume that I is stationary and ergodic with
zero mean. Let v(i) = E[&(t)&(t - i)‘] be the covariance function of I. The long-run vari-
ance of &(t), Q is then equal to c_mmY(i). Decompose Q as R, = C + A + A’, where C =

Y(0) and A = cy= 1‘-F(i). Define I? = C + A, and denote each element of Sz and r by the

@I1 012
corresponding lower case with subscripts, e.g., Ss = . Then define rz = [y12,
i 021 rn22 1
‘y2$ and let i, i‘2, p, Q,,, 6.1~~denote consistent estimators of II, I-2, p, 012, ~22. The data
transformation process in the CCR is as follows:

x;(t) = X,(t) - [i-l& +(0, Q12Q;;)~]~‘E(t),

x;(t) = X,(t) - (i-1F2)tE(t).
Simplify the expression of the data transformation process for the two-dimensional case as:

x;(f) = X,(f) + d,E(f), (3)

X2*(0 = X,(t) + Q(t), (4)

where $t and dz are real numbers. Since &(t) is stationary, XT(t) and X;(f) are cointe-
grated with the same cointegrating vector (1, -p)’ for any dt and dz. However, dt and 62
need to be seiected so that X;(t) is uncorrelated with disturbances of the regression in the
long-run.
In empirical studies, the CCR data transformation is implemented with the variable addi-
tion method (Park, 1990). The CCR estimator is then obtained by the ordinary least squares
(OLS) regression of Xl(r) on X;(t) along with appropriate deterministic terms:
Consumpt;on, Income and Cointegration 111

x;(t) = 8, + $ y,t’+ fix;(t) + E,*(t).

i= 1

The CCR is implemented with an estimated z?(f)and estimated long-run covariance pamm-
eters that are required to obtain appropriate dl and d2. An important property of the CCR
procedure is that linear restrictions can be tested by chi-square tests that are free from nui-
sance parameters. Let H(p, q) denote the standard Wald statistic to test the hypothesis y, + ,
=yp+2= . . . = y, = 0, with the estimate of variance of&(t) replaced by the long-run variance
of the CCR. In this paper, the procedure initiates with q = 1 by assuming the variables are
stochastically cointegrated. Equation (5) can then be simplified to the following form:

x;(t) = e,+ylf+ px;(t)+&,*(t). (6)

The H(0, 1) statistic test the hypothesis y1 = 0 in regression (6). Therefore, it tests the deter-
ministic cointegration restriction. When one keeps adding time polynomials into equation
(6), one will obtain equation (5). The H(0, q) statistic tests the hypothesis y1 = y2 = . . . = y,
= 0 when q time polynomial terms are added.2 The H(0, q) then tests the deterministic coin-
tegration restriction. Yet, most of the empirical studies apply the H(0, 1) test in equation (6)
to test the deterministic cointegration restriction when the variables are stochastically
cointegrated.
The cointegrating vector is known in some empirical studies. As it was shown in King et
al. (1991), the permanent income hypothesis implies that log consumption and log income
are cointegrated with the cointegrating vector (1, -1)‘. Therefore, stochastic cointegration,
which implies p = 1 and ~2 = y3 = . . . = y, = 0 in equation (5), should be tested. However,
in most cases, the cointegrating vector is unknown. Thus the H( 1, q) test for y2 = 73 = . . . =
y, = 0 alone is used to test the null hypothesis of stochastic cointegration and is a consistent
test against the alternative hypothesis of non-s~tiona~ty.
The CCR procedure requires an estimate of nuisance parameters such as long-run cova-
riance of the disturbances in the system. The VAR prewhitening method developed by
Andrews and Monahan (1992) together with the automatic bandwidth estimator developed
by Andrews (1991) are applied to estimate nuisance parameters’ in this paper. The VAR of
order one is used for prewhitening. First, a cointegrating regression based on the ordinary
least squares (OLS) is run. The OLS estimate of the cointegrating vector is used in turn to
compute the estimates of nuisance parameters for the second stage CCR. Then the second
stage CCR estimate of the cointegrating vector is used for the third stage CCR. Park and
Ogaki (199 1) recommended the third stage CCR estimate of the cointegrating vector based
on the mean-squared error (MSE) and bias when the sample size is small. In this paper, the
results for H@, q) tests are reported when the singular values for the VAR coefficient
matrix are bounded by 0.994 and the bandwidth parameter is bounded by the square root of
the sample size5 (&).

8. Testing the Null i-fypothesis of Stationarity

Park and Choi (1988) proposed the G@, q) test for the null hypothesis of stationarity for
univariate cases. It can be applied to test the stationarity of the difference of log consump-
112 HAN and OGAKI

tion and log income in this paper, based upon the assumption that the cointegrating vector
[ 1, -PI’ = [l. -11’. Let D(t) = X,(t) - X?(t), and consider the following two ordinary least
square regressions:

D(r) = ; &,t”+ S(t), (7)

k=O

D(t) = i Qk+ z(t). @I-

k=O

The G@, 9) statistic is defined as

T( 82 - 3)
Gtp,q) = &2 ’
(9)

h2 is a consistent estimator of long-run variance of the disturbance. ci is the weight of cova-

riance function with different lag truncation number 1, and is given in terms of the lag
window when the underlying process e(t) is weakly stationary. Under the null hypothesis
that D(t) is stationary (and therefore Xl(b) and X2(t) are cointegrated), the G(p, q) is chi-
square distributed with (q-p) degrees of freedom, and diverges to +m under the alternative
hypothesis of unit root non-stationarity. The G(0, q) converges to a ~92 random variable
under the hypothesis that D(t) is stationary around a level. The G( 1,q) converges to a xi _ ,
random variable under the hypothesis that D(t) is stationary around a trend. Therefore, one
can use the G(O,q) statistic to test the null hypothesis of level stationarity against the alter-
native hypothesis of difference stationarity. Specifically, one can apply the G(O,l) statistic
to test the stationarity of the difference of log consumption and log income to verify the
deterministic cointegration restriction, by assuming that the variables are stochastically
cointegrated. If log consumption and log income are cointegrated with the cointegrating vec-
tor (1, -1)’ and the deterministic cointegration restriction is satisfied by the data, the
difference of log consumption and log income should not possess any time polynomials. The
G(0, 1) test therefore can be applied to test the validity of the permanent income hypothesis.
Kwiatkowski, Phillips, Schmidt and Shin (KPSS, hereafter) (1992) presented another
test for the null hypothesis of stationa~ty for univariate cases. Again, let D(t) = X,(t) -
X*(t). They decomposed the series, D(t), t = 1,2, . . ., T, into the sum of deterministic trends,
a random walk and a stationary error:

D(f) = {t + r(t) + E(t). (I 1)

Nere r(t) is a random walk:

Consum~tion~ k~come and Cointegration 113

F-(t)= r(t - 1) + u(t), (12)

and u(r) is iid (0, c$). Because &(r) is assumed to be stationary, D(t) is stationary around a
dete~inisti~ trend under the null hypothesis that ai = 0. If 5 = 0 in equation (1 1), D(t) is
stationary around a level (ro) under the null hypothesis. KPSS considered an one-sided
Lagrange Multiplier (LM) statistic for the hypothesis a$ = 0, under the stronger assump-
tions that u(t) is normal and that &(t) is iid (0, a$). In the case of testing the null hypothesis
of trend stationarity, let e(r), t = 1,2, . . ., T, be the residuals from the regression of D(r) on
an intercept term and a time trend and i’z be the estimate of error variance from this
regression (e.g., the sum of squared residuals divided by T). The partial sum process of the
residuals is define as:

S(t) = i e(i), t = 1, 2, . . ., T. (13)

i= 1

Then the LM statistic is defined as

In the case of testing the null hypothesis of level stationarity, one can define e(r) as the
residuals from the regression of D(t) on an intercept term only. Then define the LM statistic
for the level-stationary case in the same way as described above.
We compare the test results of KPSS test without trend (or KPSS test for level stationar-
ity) with the results of G(0, 4) test; as well as the test results of KPSS test with trend (or
KPSS test for trend stationarity) with the results of G( 1, q) test in this study. The VAR pre-
whitening method developed by Andrews and Monahan (1992) together with an automatic
bandwidth estimator developed by Andrews (1991) are applied to estimate long-run vari-
ance for both G@, q) test and KPSS test.

IV. EMPIRICAL RESULTS

The consumption and income data used in this study are the same as those used by King et
al. (1991). They are quarterly observations from 1947.1 to 1988.3. The income series is
calculated by subtracting government spending from GNP. Then real consumption and real
income are divided by total population at the end of the month at each quarter to obtain per
capita values.
Table 1 presents the CCR test results. Either log consumption or log income can be used
as regressand in the CCR.6 When log consumption is used as regressand, the p-value of
H(0, 1) is 0.577. Under the assumption that the variables are stochastically cointegrated,
the deterministic cointegration restriction cannot be rejected in terms of the H(0, 1) test.
The result is robust to the choice of regressand. Next, we test stochastic cointegration and
find that there is no evidence against it by the H( 1, q) test for q = 2,4,7. The results are con-
sistent with findings of King et al. (1991). They tested the null hypothesis that the data are
integrated but not cointegrated using procedures developed by Johansen (1988, 1991) and
114 HAN and OCAKI

Tabk I. Canonical Cointegrating Regression Results

Log consumption 1.048 0.311 0.120 5.663 8.047

(0.016) (0.577) (0.729) (0.129) (0.235)
Log GNP 0.953 1.810 0.022 1.250 3.062
(0.024) (0.179) (0.88 1) (0.741) (0.801)

Nora; Andrew and Monahan’s (1992) VAR prewhitening method together with the Andrew f 1991) automatic bandwidth esti-
mator are used.
a Standard errors are shown in parentheses.
b p-values are shown in parentheses. This statistic tests the deterministic cointegration restriction.
Cp-values are shown in parentheses. These statistics test the null hypothesis of stochastic cointegration.

Table 2. Tests for the Null Hypothesis of Level Stationarity

GV, I)” G(0,21a Gf 0,3? KPSSb

3.399 3.453 4.829 0.30029

(0.065) (0.178) (0.185)

Notes: Andrew and Monaban’s (1992) VAR pre~~hi~enin~ method together with the Andrew’ (1991)
automatic bandwtdth estimator are used to calculate the long-run variance.
Ap-values are shown in parentheses.
b According to Kwiatkowski et al. (p. 166, 1992), the upper 5% and 10% critical values for this
statistic are 0.463 and 0.347, respectively.

Stock and Watson ( 1988).7 On the contrary, the Htp, q) tests the null hypothesis of cointe-
gration. Therefore, it is difficult to make a comparison between the H@, q) test and the
Johansen’s test. King et al. (1991) left the constant term unrestricted in the application of
Johansen’s procedure, they implicitly assumed that the cointegrating vector eliminates
both deterministic trends and stochastic trends, therefore, their test could have power
against the deterministic cointegration restriction as well as stochastic cointegration.
Table 2 presents the test results on deterministic cointegration of the difference of log
consumption and log income from Park and Choi’s G(0, q) test and KPSS test for level
stationarity. First, it is found that the null hypothesis of level stationarity cannot be
rejected by G(0, I), G(0, 2), and G(0, 3) tests. Second, the KPSS test statistic for the
level-stationary case is only 0.30029. According to Kwiatkowski et al. (p. 166, 1992), the
upper 5% and 10% critical values for this statistic are 0.463 and 0.347, respectively. The
null hypothesis of level stationa~ty cannot be rejected by KPSS test either. The differ-
ence of log consumption and log income is stationary around a level. Therefore, log
consumption and log income are cointegrated with the cointegrating vector (1, -1)’ and
the deterministic cointegration restriction is satisfied. The results are consistent with the
CCR test results.
We also use the Gf 1, q) test and KPSS test with trend to verify stochastic cointegration
of the difference of log consumption and log income. The results are presented in Table 3.
It is found that all G(l, 2), G(1, 3), and G(l, 4) tests cannot reject the null hypothesis of
trend stationarity. The KPSS test statistic for the trend-stationary case is 0.038. According
to Kwiatkowski et al. (p. 166, 1992), the upper 10% critical value for this statistic is 0.119.
The null hypothesis of trend stationarity cannot be rejected by KPSS test. Both the G( 1, q)
Consumption, income and Cointegration 115

Table 3. Tests for the Null Hypothesis of Trend Stationarity

G(l, 2Ja G(1,3)" WI, 4? KPSsb

0.099 2.65
1 2.652 0.038
(0.753) (0.266) (0.449)

Notes:
Andrew and Monahan’s (I 992) VAR prewhitening method together with the Andrew’ (I 991)
automatic bandwidth estimator are used to calculate tbe long-run variance.
ap-values are shown in parentheses.
b According to Kwiatkowski et al. (p. 166, 1992). the upper 10% critical values for this statistic
is 0.119.

test and KPSS test with trend have consistent results on stochastic cointegration of the dif-
ference of log consumption and log income.
Han (1996) reported small sample properties of the CCR tests. It was found that the CCR
tests perform reasonably well in terms of size and power when sample size is small.
Skanderson (1994) also performed Monte Carlo simulation to compare the G(p, q) test with
KPSS test. It was found that these two tests perform well and are very similar in terms of
size and power for small samples. The empirical testing results in this study therefore are
reliable even when the sample size is relatively small.

V. CONCLUSIONS

This paper shows that the United States post war quarterly consumption and income data
support the joint hypothesis of the difference stationarity and a version of the permanent
income hypothesis. The deterministic cointegration restriction cannot be rejected by the
CCR tests, the G(0, q) test, and KPSS test without trend. The results are consistent with the
finding by Kuznet (1946) that the saving rate is stable in the United States in the long-run.

ACKNOWLEDGMENTS

The authors would like to thank Robert King for providing some of the data used in this
paper and for his helpful comments. We would also like to thank the editor and an anony-
mous referee for their helpful suggestions. All remaining errors are ours. The second
author gratefully acknowledges financial support by the National Science Fundation Grant
No. SES-9213930.

NOTES

1. Neusser (1991) confirmed this finding for some other industrial countries, as well
as for the United States by using slightly different data set. Both King et al. (1991) and
Neusser (199 1) included other economic variables such as investment to investigate a neo-
classical model that includes the PIH. The present paper focuses on the PIH.
116 HAN and OCAKI

2. The choice of 9 should be relatively small, according to Han (1996) which examines
the small sample properties of the CCR.
3. Automatic bandwidth estimator by Andrews (1991), ST, is constructed from fitting
AR( 1) to each disturbance.
4. This follows the method in Andrews and Monahan (1992).
5. These parameters are not bounded for the seocnd stage CCR and the third stage
CCR, because it is found that this would often deteriorate the MSE and bias in our Monte
Carlo simulations.
6. We use Ogaki’s (1993) GAUSS CCR package for the empirical work. The package
is available from the second author on request.
7. King et al. (1991) considered consumption, income and interest rate in their three-
variable model. Their Table 1 Panel B presents formal tests for the null hypothesis that the
three variables are not cointegrated so that there are three unit roots in the companion
matrix. Their testing results are consistent with the one-unit-root specification. Therefore,
they concluded that there is one common trend and the variables are cointegrated.

REFERENCES

Andrews, D.W.K. (1991). Heteroskedasticity and autocorrelation consistent covariance

matrix estimation. Econometrica, 59, 817-858.
Andrews, D.W.K., & Monahan, C.J. (1992). An improved heteroskedasticity and
autocorrelation consistent covariance matrix estimator. Econometrica, 60, 953-966.
Campbell, J.Y. (1987). Does saving anticipate decline labor income? An alternative test of
the permanent income hypothesis. Econometrica, 5.5, 1249-1273.
Campbell, J.Y., & Perron, P. (1991). Pi@alls and opportunities: What macroeconomists
should know about unit roots. NBER Macroeconomics Annual, Cambridge, MA:
MIT Press.
Christiano, L.J., & Eichenbaum, M. (1989). Unit roots in real GNP-Do we know and do
we care? Paper presented at the Carnegie-Rochester Public Policy Conference, April
21-22,1989.
Cochrane, J.H. (1988). How big is the random walk in GNP? Joarna~ of Pol~ticul
Economy, 96, 893-920.
Cochrane, J.H., & Sbordone, A.M. (1988). Multivariate estimates of the permanent
components of GNP and stock prices. Journal of Economic Dynamics and Control,
12, 255-296.
Cooley, T.F., & Ogaki, M. (1996). A time series analysis of real wages, consumption, and
asset returns: A cointegration-Euler equation approach. Journal of Applied
Econometrics, II, 38-71.
Engle, R.F., & Granger, C.W.J. (1987). Co-integration and error correction:
Representation, estimation, and testing. Econometrica, 55, 25 l-276.
Han, H-L (1996). Small sample properties of canonical cointegrating regressions.
Empirical Economics, 21, 235-253.
Hansen, B.E. (1992). Efficient estimation of cointegration vectors in the presence of
deterministic trends. Journal of Econometrics, 53, 87-121.
Consu~~tjon, income and Cojntegra~ion 117

Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of Economic

Dynamics and Control, 12, 23 1-254.
Johansen, S. (1991). Estimation and hypothesis testing of cointegrating vectors in Gaussian
vector autoregressive models. E~onametricu, 59, 155 l- 1580.
King, R.G., Plosser, C.I., Stock, J.H., & Watson, M.W. (1991). Stochastic trends and
economic fluctuations. American Economic Review, 81, 819-840.
Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., & Shin, Y. (1992). Testing the null
hypothesis of stationarity against the alternative of a unit root: How sure are we that
economic time series have a unit root. Journal of Econometrics, 54, 159-178.
Kuznet, S. (1946). rational income: A surn~na~ of~nd~ngs. National Bureau of Economic
Research, New York.
Neusser, K. (1991). Testing the long-run implications of the neoclassical model. Journal of
Monetary Economics, 27, 3-37.
Ogaki, M. (1992). Engel’s law and cointegration. Journal of Political Economy, 100,
1027-1046.
Ogaki, M. (1993). CCR: A user guide. Rochester Center for Economic Research Working
Paper, 349, University of Rochester.
Ogaki, M., & Park, J.Y. (1996). A cointegration approach to estimate preference
parameters. Journal ofEconometrics, forthcoming.
Park, J.Y. (1990). Testing for unit roots and cointegration by variable addition. Advances
in Econometrics, 8, 107-133.
Park, J.Y. (1992). Canonical cointegrating regressions. Econometrica, 60, I 19-143.
Park, J.Y., & Choi, B. (1988). A new approach to testing for a unit root. CEA Working
Paper, 88-23, Cornell University.
Park, J.Y., & Ogaki, M. (1991). Inference in cointegrated models using VAR prewhitening
to estimate short-run dynamics. Rochester Center for Economic Research Working
Paper, 281, University of Rochester.
Phillips, P.C.B., & Ouliaris, S. (1988). Testing for cointegration using principal
components methods. Journal of Economic Dynamics and Control, 12, 205-230.
Skanderson, D. (1994). Evidence on the Fisher equation: The long-run and the short-run.
Ph.D. Dissertation, University of Rochester.
Stock, J.M., & Watson, M.W. (1988). Testing for common trends. Journal of the American
statistical Association, 83, 1097-I 107.
West, K.D. (1988). Asymptotic normality when regressors have a unit root. Econometrica,
56, 1397-1417.